Capabilities

@llm-ports/capabilities ships seven cognitive operation factories. Each factory takes config (port, schema, prompt fragments, hooks) at definition time and returns a typed function you call per-input.

The seven (extracted from BEPA's production stack):

Factory	Returns	Default temperature	Use for
createClassifier	typed object (Zod)	0	Pick one of N categories
createScorer	typed object (Zod)	0.1	Rate against a rubric
createExtractor	typed object (Zod)	0	Pull structured fields
createSummarizer	text	0.2	Compress meaning-preserving
createDrafter	text	0.4	Generate text in a persona
createPlanner	typed object (Zod)	0.2	Decompose into steps
createAnalyzer	typed object (Zod)	0.3	Evaluate / critique / compare

10 more capabilities ship in v0.2: tag, detect, expand, rewrite, redact, respond, decide, answer, rerank, agent.

Why factories?

The factory pattern is the most important API choice in llm-ports. Three alternatives were considered and rejected:

• Plain helpers: force re-passing config (rubric, schema, hooks) on every call. Boilerplate at the call site.
• Decorators: require experimentalDecorators config and pollute the type system.
• Task descriptors: separate definition from execution, which is the wrong cut for LLM work where prompt fragments ARE the definition.

Factories let you bind rubric, schema, boundaryExamples, systemContext, and hooks once at definition time. Reading the call site shows what is bound and what is varying.

Common shape

Every factory follows the same template:

const fn = createX({
  // Required
  port: llm,
  schema: z.object({...}),       // for structured-output capabilities
  schemaName: "operation-name",  // appears in observability events

  // Prompt fragments (string OR async function)
  rubric: "...",                  // per-capability variant of the framework
  examples: "...",
  systemContext: async (input) => `dynamic ${await lookup()}`,

  // Routing
  taskType: "triage",             // matches LLM_TASK_ROUTE_TRIAGE in env
  priority: 2,
  temperature: 0.2,
  maxOutputTokens: 1024,

  // Hooks (errors caught and logged but never re-thrown)
  onBeforeCall: async (input) => { /* ... */ },
  onResult: async (event) => { /* ... */ },
  onError: async (err, input) => { /* ... */ },
});

Then call with input:

const result = await fn({ content: "...", contextOverride: "per-call extra context" });

Hooks and the CapabilityEvent

Every successful call invokes onResult with a standardized CapabilityEvent:

interface CapabilityEvent<TOutput> {
  capability: string;             // "classify" | "score" | ... — the factory's identity
  schemaName: string;             // the user-supplied operation name
  modelId: string;                // which model actually ran
  providerAlias: string;          // which provider alias was selected
  usage: { inputTokens, outputTokens, totalTokens };
  cost:  { inputUSD, outputUSD, totalUSD };
  latencyMs: number;
  output: TOutput;                // the validated typed result
  validationAttempts?: number;    // present for structured-output capabilities
}

Wire this to your analytics pipeline directly:

const classify = createClassifier({
  port: llm,
  schema: TriageSchema,
  schemaName: "email-triage",
  onResult: async (event) => {
    await myDB.insert("llm_events", {
      capability: event.capability,
      schemaName: event.schemaName,
      modelId: event.modelId,
      cost: event.cost.totalUSD,
      latencyMs: event.latencyMs,
      timestamp: new Date(),
    });
  },
});

The standard event shape across all capabilities means you can write one analytics handler and reuse it.

Dynamic prompt fragments

Prompt fragments accept either a string or a function returning a string. Functions can be sync or async, enabling DB lookups, feature flags, or context-derived content:

const classifyEmail = createClassifier({
  port: llm,
  schema: TriageSchema,
  schemaName: "email-triage",
  rubric: async () => await loadRubricFromConfigService(),
  systemContext: async (input) => {
    const sender = await getSenderProfile(input.content);
    return `Sender warmth: ${sender.warmth}, last interaction: ${sender.lastSeen}`;
  },
});

Resolution happens lazily, per-input. No pre-compilation, no caching unless you add it yourself.

Hook safety

Hooks should not break the call. If your onResult handler throws, the capability:

1. Catches the error
2. Logs it via console.warn
3. Returns the model's output to the caller anyway

This is intentional. Observability hooks are nice-to-have; production calls must succeed even when telemetry is broken.

If you want hard-failure on hook errors, throw in onError instead — that one re-throws.

Reading next

• createClassifier → — most-used capability
• createDrafter → — text generation in a persona
• Source code: packages/capabilities/src/